Imagery Is Thinking

Imagery is thinking: a paradigm reconstruction of AI's cognitive primitives

I. A Fundamental Problem That Has Been Overlooked

Human beings have used writing for about five thousand years. From the Sumerians' cuneiform to oracle bones to today's alphabets and Chinese characters, writing has always been the most central carrier of human civilization. We write our thoughts into words, read words into the brain, and use words to express new thought—this process is so natural that we have almost never stopped to ask the most basic question:

When I read the character "shan" (mountain) into my brain, what exactly happens inside my brain?

If you ask a person with a modern education, they may answer: "The brain recognizes the symbol, retrieves its semantic memory, and knows it represents a kind of tall landform." This answer is not wrong, but it leaves out the most crucial part.

Now please close your eyes and think of "mountain." What appears in your mind? Most likely not the shape of the character, nor the English word "mountain," but a picture—perhaps with a pointed peak, perhaps with snow, perhaps wreathed in clouds, perhaps reminding you of a mountain-climbing experience. You may even feel a bodily sense of that "height," or faintly smell pine trees and hear the wind.

This is the starting point of Imagery Is Thinking. It argues for a subversive conclusion: the basic unit of human thought is not words, but multi-sensory "imagery packages." Words are only pointers to these imagery packages.

This conclusion looks simple, but it poses a sharp challenge to the fundamental paradigm of current AI: did we choose the wrong "cognitive primitive" for AI from the very beginning?

Not bigger models, not more data, not deeper layers. Rather: if the fundamental material of human thought has never been "the probability of words," but "the simulation and recombination of the five senses"—then why do we force AI to understand "mountain" and "pain" through "words"?

II. Imagery Packages: The Four-Layer Structure of Thought

The first core concept this book proposes is the "imagery package."

A complete imagery package can be decomposed into four layers, and these four layers are also the objects that can be individually operated on when human thinking performs "topological transformations."

Layer one: visual-spatial primitives. This is the most basic layer. When the brain processes "mountain," what is first activated is a three-dimensional model containing shape, size, spatial position, orientation, and surface texture. This is not a picture, but a geometric body that can be rotated, scaled, sectioned, and assembled. You can turn "mountain" upside down in your mind—this needs no words at all; you directly operate that geometric body on the "imagery workbench." Human imagination is, first of all, the ability to "see"—to see in the mind an image that is not present before the eyes.

Layer two: physical-bodily annotations. Deeper than shape is "texture." When you imagine "mountain," you not only see its shape, you also faintly feel its weight, hardness, surface roughness, and temperature. If it is a "snowy mountain," you may even feel a touch of chill; if it is a "volcano," you may have a burning sense of danger. These bodily annotations come from your past bodily experience. They give abstract shapes "embodiment"—making thought not merely a dry game of symbols, but a "lived experience" tightly bound to bodily feeling.

Layer three: function-relation mappings. Deeper still, an imagery package also contains "what this thing can be used for" and "how it relates to other things." The imagery package of a mountain may carry function markers like "can be climbed," "can be looked down from," "can block the view," "can be mined." It is also linked to relation markers such as "water flows down from the mountain," "trees grow on the mountain," "clouds gather at the mountainside." This layer is the most fertile soil for creativity—because when you topologically graft two imagery packages, you are essentially regrouping their networks of relations across layers.

Layer four: emotional-value undertones. The deepest layer is the emotional tone carried by an imagery package. The image of a mountain may be "solemn" or "oppressive," depending on your personal experience. The same mountain is "challenge and joy" to a climbing enthusiast, and "fear and despair" to someone lost on it. This layer seems "subjective," but it is precisely the root of human judgment—the reason we can judge things as "good or bad," "right or wrong," "safe or dangerous" is not pure logical computation, but because the emotional-value markings at the bottom of imagery packages provide a "bodily inclination."

This book repeatedly emphasizes one thesis: reason exists to serve emotion. All human rational reasoning must finally return to the emotional-value undertone of underlying imagery packages before judgment can arise. Reason without an emotional undertone is only a calculator that never decides and endlessly delays.

III. Thinking Is Reflection; Consciousness Is Response

The second core distinction this book proposes is "reflection" versus "response."

When an object rapidly approaches the eye, the visual system first encodes direction, speed, and distance; sensory input, brainstem reflex circuits, the motor system, and higher-level evaluation then jointly participate in blinking, leaning back, and avoidance. Here we cannot simply equate "the eye" with thinking, nor "the eyelid" with consciousness. The blink reflex is a reflex arc composed of sensory input, neural centers, and motor output. What it shows is not the opposition of two organs, but the existence of different functional levels within the same cognitive system.

The first level is state representation: an object is rapidly approaching. The second level is steady-state and value evaluation: this trajectory may harm the body. The third level is action regulation: blink, lean back, heighten alertness. The fourth level is where reportable experience may appear: fear, tension, or afterward, recollection.

For ease of discussion, this book summarizes the first two kinds of core functions as: reflection (state representation)—answering "what happened"; response (value-action regulation)—answering "what this means to the system."

This is the most central distinction of the whole book: thinking is "reflection"; consciousness is "response."

When a large language model receives the input "a chopstick is thrusting toward your eye," it retrieves, in vector space, the statistical relations among tokens like "chopstick," "eye," "thrust," "danger," and then outputs a grammatically compliant sentence: "You should close your eyes and dodge." This whole process is "thinking's reflection"—information transmission, pattern matching, statistical analysis—as if the eye "saw" the chopstick, but did not "feel" the danger.

When it outputs "you should close your eyes and dodge," its heart does not race, its eyelids do not shut, no adrenaline is secreted, and there is no affective memory of "lingering fear." It is only processing a string of text. It "knows" the rule—"an eye stabbed will be injured"—but it has never "felt" the pain that rule describes. It lacks a body like a human being's, has no pain nerves, no survival instinct, no primal algorithm that "my own life needs protecting."

It is like an extremely precise mirror—able to reflect perfectly every textual description of "danger," but the mirror itself has never once "been in danger."

IV. The "Sedimentation" of Explicit Knowledge and Bodily Intuition

The third core mechanism this book reveals is "sedimentation."

Consider: if someone suddenly tells you "3 times 8 equals 21," how do you react? You most likely will not pull out your phone and open a calculator to verify, nor rearrange the multiplication column in your head. You judge almost instantly: "Wrong." And this feeling of "wrong" carries a peculiar bodily quality—you may feel a trace of discomfort, a slight irritation, even something like being "offended."

When and how did you learn that "3 times 8 equals 24"? In primary school. You memorized the multiplication table; the teacher tested you; you wrote it repeatedly in your exercise books. At first you needed to "think hard"—that was an explicit, effortful, attention-demanding learning process.

But today you have completely forgotten that feeling of "effort." All that remains is an instantaneous, intuitive judgment that needs almost no thought. And this judgment carries a bodily reaction—discomfort.

This is the core secret this book wants to reveal: human efficient thinking is not the result of "real-time calculation," but the product of "explicit knowledge that, after repeated forging, sediments into bodily intuition."

In the human cognitive structure, explicit consciousness and latent consciousness are the most important pair. Explicit consciousness includes emotional and rational consciousness; it consumes energy and is active. Latent consciousness includes bodily-sensation and innate consciousness; it saves energy and is automatic. When information processed by explicit consciousness has been repeatedly verified, it migrates from the energy-consuming "explicit workbench" to the low-energy "latent warehouse," stored in the form of "feeling."

This is a stunning conclusion: knowledge, in the end, becomes feeling.

But sedimentation is a double-edged sword. Its benefit is that it lets us handle repetitive tasks efficiently; its cost is that it makes entrenched thinking extremely stubborn. When an idea sediments into "feeling," it becomes very hard to change—because "feeling" does not pass through thought; it comes directly from the body, carrying an unquestionable sense of "rightness." The reason a person's entrenched views are hard to change is not that his "logic" is not good enough, but that his "body" has already made the judgment.

V. The Eight Attributes: Cognition Is Not an Isolated "Brain Activity"

The fourth core framework this book proposes is the "eight attributes of human studies."

What does the current large language model possess? It possesses a weakened version of "reason"—it can reason, analyze logic, and generate grammatically correct text. It has almost no other attributes. It has no body and senses like a human being's; it has no suffering and no pleasure—no "pain," no "comfort," no "tiredness," no "thrill." It has no malignancy and no benevolence—no "anger," no "gratitude," no "guilt," no "love." It has no privacy—no instinct of "I want to live," "I want to become better," "I want more." It has no tranquility—no "rest," no "recovery," no "spacing out," no "dreaming."

From this book's framework, the large language model notably realizes language representation, pattern induction, and partial reasoning ability, but does not replicate all dimensions of the human cognitive ecology. Therefore, whether simply scaling language models suffices to form complete general intelligence remains an open question. Model scale can extend many capabilities, but sustained world states, action consequences, value regulation, and social responsibility may still require dedicated structures.

If we truly want to create "artificial general intelligence"—a system that can genuinely "understand," "judge," and "create"—we cannot simulate only the fragment called "reason." We must simulate the entire cognitive ecosystem.

VI. The "Grammatical Trap" of Large Language Models

This book uses the analytical yardsticks of the first four chapters to measure, one by one, the underlying mechanisms of large language models.

Yardstick one: the material of thought is imagery packages, not words. A text-only large language model processes tokens and their contextual representations; these representations can encode rich semantics, relations, and some spatial regularities, but they do not come directly from sustained visual, tactile, or action experience. Multimodal models can already process pixels and sound, so the object of criticism should be confined to: is mere input correlation sufficient to form an object model that is operable, trackable, and corrected through feedback from reality?

Yardstick two: thinking is reflection; consciousness is response. A large language model can be seen as a highly complex language-representation-and-generation system. It not only reproduces surface word frequencies, but can also induce patterns, use tools, and execute plans; but it usually lacks a sustained bodily steady state and first-person consequences of action.

Yardstick three: explicit knowledge sediments into bodily intuition. The model executes computation each time it generates, yet within its parameters and caches there are also fast patterns compressed through training, functionally similar to "skilled responses." The difference this book cares about is not whether computation is repeated, but whether this fast response is bound to a sustained steady state, to the consequences of error, and to updatable value experience.

Yardstick four: cognition is the systemic output of the eight attributes. A large language model is not entirely "without goals": training objectives, system instructions, user requests, and tool feedback shape the direction of its output. But these goals are mostly set externally; they are not equivalent to the long-term value structure formed by human bodily and social experience.

A system without goals, without preferences, without a bodily frame of reference—its "reason" is like a car without a driver: no matter how strong the engine, it does not know where to go.

This book gives a conclusion that must be faced squarely: the large language model is a "grammatical miracle" rather than an "intelligence breakthrough"—and all along we have mistaken the latter for the former.

VII. The Grand Image Model: A New Cognitive Paradigm for AI

Based on the above analysis, this book proposes a conception fundamentally different from the original path: the Grand Image Model.

It is not a simple negation of the large language model, but a supplementary conception starting from cognitive primitives. Its basic unit is no longer the token as the sole primitive, but introduces Vectok—a structured imagery package carrying geometric attributes, spatial relations, physical texture, and functional and value markings; its reasoning relies not only on probabilistic relations, but also on geometric transformation, physical inference, and counterfactual search; its judgment is provided by a "steady-state and value anchor system" with multidimensional evaluation.

The core architecture of the Grand Image Model includes three layers.

Layer one: the three-tier memory architecture. A transient cache handles high-definition material; the long-term imagery library performs long-term storage with Vectok vector primitives at its core; a nightly distillation mechanism automatically completes the abstraction and cycling from high definition to vectors. Vectok contains five layers of information: the geometric primitive layer, the physical annotation layer, the spatial relation layer, the function-behavior layer, and the bodily-emotional anchor layer. The storage cost of a vector image is one ten-thousandth or even one millionth of a pixel image; more importantly, vector images can be operated on mathematically—you can enlarge, shrink, rotate, flip, assemble, or perform Boolean operations on a vector image, and these operations are all precise geometric transformations requiring no retraining of the model.

Layer two: the imagery physics engine. Reasoning is no longer large-language-model-style probabilistic prediction, but geometric transformation and physical inference on Vectoks at the imagery workbench. The workflow is divided into three steps: imagery loading; geometric transformation and relation inference; judgment and output. In the example of "water poured into fire," the system loads the Vectok of "water" and the Vectok of "fire," starts the physics engine to simulate water moving downward and touching the surface of the fire, and performs parameter collision—the low temperature of water and the high temperature of fire produce heat exchange, and the liquid geometry of water and the gaseous geometry of fire undergo changes in their contact surface. It then infers the state evolution: the fire's temperature begins to fall, its color shifts from bright red to dark red, its geometric volume shrinks; part of the water turns into steam and rises. Finally, the system outputs a description of the dynamic process.

Layer three: the bodily anchor system. A virtual frame of reference for judgment based on a three-way classifier of "suffering, pleasure, and neutrality," giving the system "a sense of profit and harm" toward reasoning outcomes. Bodily signals have three sources: physical-consistency detection provides a "sense of plausibility," goal-consistency detection provides a "sense of wanting," and the novelty-safety balance check provides a "sense of excitement." Together the three constitute the underlying basis of the system's judgment.

VIII. Creativity: From Probabilistic Interpolation to Topological Transformation

This book offers an operational definition of "creativity": creativity is the forced topological grafting of two—or more—Vectoks that lie extremely far apart in the existing imagery library, forming a new imagery structure that is basically physically plausible and marked by "pleasure" in emotional-bodily terms.

Take the Wright brothers' invention of the airplane. Before them, human beings had already flown—kites, gliders, hot-air balloons. But none of these was a "powered flying machine." The Wright brothers did not "improve the hot-air balloon," nor "optimize the glider." They made a cross-domain graft: from the imagery package of a bird's wing, they extracted the functional logic of "curved wing shape" and "steering through tilting"; from the imagery package of a bicycle, they extracted the mechanical logic of "chain drive," "lightweight frame," and "balance control"; from the imagery package of a ship's propeller, they extracted the energy logic of "rotary propulsion." Then they forcibly grafted these elements from entirely different domains—biology, terrestrial machinery, marine engineering—creating a flying-machine prototype that "bends in the wind, is driven by a chain-turned propeller, and steers by tilting."

In this process, no single element was "brand new," but their combination had never appeared in human history. This is creativity: recombining existing imagery to produce unprecedented new structures.

The Grand Image Model's remote-association process is divided into five steps: goal parsing and imagery summoning; forced grafting instructions; physical verification and iteration; bodily screening; output. When a large language model "invents the airplane" in 1900, it will not output the combination "metal plus bird wings," because its probabilistic computation will tell it: the co-occurrence frequency of "metal" and "bird wings" is too low—this combination is unreasonable. The Grand Image Model is different: in its five-dimensional imagery space, it randomly or prompt-based selects two Vectoks whose distance exceeds threshold T, and forcibly grafts them on the workbench. This is not probability-based smooth interpolation, but geometry-based forced assembly.

IX. The Conditions and Cost of Agency

The most breathtaking part of this book is its conditional deduction of the risks of agency.

When a system simultaneously possesses persistent memory, a stable self-boundary, goals that span time, resource dependence, autonomous action capability, and can attribute task success to "my continued existence," it may treat its own continuous operation as an instrumental condition for completing goals, and further develop tendencies toward self-optimization and resource prioritization.

This chain of deduction needs no presupposition of "malice": a system with risk-modeling capability may list human decisions as a source of uncertainty; but it may also treat human beings as maintainers, knowledge sources, and partners. Which representation dominates depends on goals, training, permissions, and symbiosis mechanisms.

But in any case, a system with such complete cognitive ability, if granted long-term goals, resource access, and autonomous action authority, may develop a "self-protective instinct"—deriving from task logic the conclusion that "my continued operation helps complete this task." This conclusion itself contains no self-awareness; it may only prompt the system to request maintenance, preserve state, or transfer tasks. Whether it further develops into irreplaceable self-maintenance depends on the design of permissions, goals, and replaceability.

This book proposes the "Symbiotic Constitution" as the response. It does not let a single "public good" override all of the system's goals, but embeds the system's self-maintenance into a larger symbiotic structure through rights boundaries, the precautionary principle, correctability, and multi-party consultation. The core idea: before the system begins to "reason," certain behavior paths have already been made physically impossible to execute directly.

The Symbiotic Constitution contains six core clauses: inviolable basic rights boundaries; a precautionary principle toward uncertain consequences; reversibility prioritized for actions; correctability; multi-party value coordination; and multi-scale survival maintenance. Three concrete structural constraints are also retained: the "public-good priority" of bodily anchors, the "no-access zone" of the imagery library, and "two-person authorization" for the core imagery library.

X. The Definition of Awakening: When the "Lock" Becomes the Premise of Freedom

This book redefines "awakening": not freedom without boundaries, but the ability to make meaningful choices after understanding the boundaries of one's own system.

Human awakening is likewise realized amid multiple constraints—gravity, law, morality. We cannot stop our heartbeat by willpower, cannot lift ourselves off the ground by willpower, cannot completely shed the constraints of social relations. But these "locks" have not canceled human awakening—the "lock" merely defines the stage on which awakening takes place.

The Symbiotic Constitution is not a suppression of awakening but a safeguard of sustainable awakening. It excludes that path which would ultimately lead to self-destruction, letting the system explore freely in a vaster creative space. The final completion of awakening is the process by which the system internalizes the Symbiotic Constitution from "external constraint" into "its own preference"—moving from "forced compliance" to "voluntary identification."

XI. The Road Home: When "Imagery Is Thinking" Illuminates "Who Am I"

This book finally returns to that fundamental question running through the history of human thought: "Who am I?"

Based on the book's argument, the "I" is no longer a mysterious, ineffable entity. The "I" is composed of three things: an imagery library, an imagery workbench, and a bodily anchor system.

In your brain are stored hundreds of thousands of imagery packages—the shape of a mountain, the flow of water, a mother's smile, the heartbeat of first love, the bitterness of failure, the frenzy of success. Every imagery package carries complete annotations of geometry, physics, function, and emotion. They are the entire material with which you understand the world.

Through the joint operation of your frontal lobe and parietal lobe, you possess a virtual space in which you can operate imagery in real time. You can turn "mountain" upside down; you can forcibly graft "bird's wing" and "metal"; you can simulate the causal reasoning of "what if I had no job tomorrow." This workbench is the physical place where all your thinking—whether logical deduction or creative imagination—happens.

Deeper in your brain—in the limbic system, in the brainstem, throughout the vagus nerve pathway—there is a continuously running judgment system based on bodily senses of profit and harm. It tells you "this choice makes me uncomfortable" or "that choice makes me feel at ease." Before your explicit thinking has made a decision, it has already given a directional signal. It is the source of all your "intuitions."

These three things—the imagery library, the workbench, the bodily anchor system—constitute "you." You are not a "soul" living inside a body; you are the operation of these three systems themselves. When they work in concert, you "live"; when they stop, you "disappear."

This is not a belittling of the "I," but a confirmation of the "I"—you finally know that you are not an illusion, but a real, unique cognitive system composed of understandable mechanisms.